national energy research scientific computing center (nersc) · incite program at nersc 2006 •...

National Energy Research Scientific Computing Center (NERSC)

Jonathan Carter

SOS 11June 12, 2007

NERSC Mission

The mission of the National Energy Research Scientific Computing Center (NERSC) is to accelerate the pace of scientific discovery by providing high performance computing, information, data, and communications services for research sponsored by the DOE Office of Science (SC).

Science-Driven Computing Strategy 2006 -2010

Science-Driven Computing

National Energy Research Scientific Computer Center (NERSC)

– computational facility for open science

– international user community

– 2500 users– 300 projects

NERSC is enabling new science

Archiving for Genomics Research• Production Genome Facility (PGF) at

Joint Genome Institute (JGI) is producing sequence data at increasing rate– 2 million files per month of trace

data (25 to 100 KB each) – 100 assembled projects per month

(50 MB to 250 MB)– several very large assembled

projects per year (~50 GB). – total about 2 TB per month on

average• NERSC and PGF staff collaborated to

set up data pipeline using ESnet’snew Bay Area MAN

• NERSC provides scientific data archive capability

NERSC User George Smoot wins 2006 Nobel Prize in Physics

Mather and Smoot 1992

COBE Experiment showed anisotropy of CMB

Cosmic Microwave Background Radiation (CMB): an image of the universe at 400,000 years

CMB Computing at NERSC

• CMB data analysis presents a significant and growing computational challenge, requiring– well-controlled approximate algorithms– efficient massively parallel implementations– long-term access to the best HPC resources

• DOE/NERSC has become the leading HPC facility in the world for CMB data analysis– about 1,000,000 CPU-hours/year for the last 4

years– 6.25 TB project disk space– 300 TB HPSS data– about 10 experiments and 100 users

CMB is Characteristic for Large-Scale Projects at NERSC

• Petaflop/s and beyond computing requirements• Algorithm and software requirements• Use of new technology, e.g. NGF• Service to a large international community• Exciting science

INCITE Program at NERSC 2006

• Precision Cosmology Using the Lyman Alpha Forest – Mike Norman, SDSC

– Increase understanding of the dark energy and dark matter

– Large memory requirements needed 256 seaborg nodes, 10TB disk space

– Generated > 60 TB data – Systems group and consulting staff

shepherd through large debug jobs to track down memory consumption of code

• Particle-in-Cell Simulation of Laser Wakefield Particle Acceleration – Cameron Geddes, LBNL

– Produce detailed 3D models of laser-driven wakefield particle accelerators

– 4.6 million hours devoted to full 3D higher resolution simulations

– Clarify mechanisms for beam formation and evolution

• Reactions of lithium carbenoids, lithium enolates, and mixed aggregates – Larry Pratt, Fisk University

– Investigate the structure and reactions of organolithium compounds

– Simulations involving electron correlation change equilibrium geometries significantly

– Fine-grained parallelism requires powerful large-memory SMP nodes.

“I hope we can continue to work with NERSC in the future. I think your center did a

fabulous job in supporting us. ” Robert

Harkness

Impact on Science Mission

Impact on Science Mission

• Majority of great science in SC is done with medium-to large-scale resources

• In 2006, NERSC users reported the publication of 1437 papers that were based wholly or partly on work done at NERSC (http://www.nersc.gov/news/reports/ERCAPpubs06.php)

Science-Driven Services

Science-Driven Services

• Provide the entire range of services from high-quality operations to direct scientific support

• Enable a broad range of scientists to effectively use NERSC in their research

• Concentrate on resources for scaling to large numbers of processors/cores, and for supporting multidisciplinary computational science teams

Science-Driven Services

• Consulting– One-on-one code tuning– Debugging– Software installation– Data manipulation advice: bbcp, gridftp, NGF, etc.

• Systems– Queue priority and time limits– Increased disk quotas

• Analytics– Large (INCITE/SciDAC) projects tend to produce the most output

and often require large input datasets• Networking

– Network tuning, software installs (bbcp) to boost transfer bandwidth between NERSC and ORNL from 6.4 MB/s to 24 MB/s for combustion project

– >60 TB transferred to SDSC using gridFTP at 25-55 MB/s for astrophysics project

– Tuned network connections and replaced scp with hsi: transfer rate increased from 0.5 to 70 MB/s for astrophysics project

Example of Special Assistance

Photosynthesis Project PI: William Lester, UC Berkeley

• MPI tuning: 15-40% less MPI time

• Load balancing: scaling from 256 to 4,096 procs

• More efficient algorithm for random walk procedure

• Wrote parallel HDF5 I/O layer

“We have benefited enormously from the

support of NERSC staff.”

Thermonuclear Supernovae Project

PI: Tomasz Plewa, U. Chicago• Resolved problems with

large I/O by switching to a 64-bit environment

• Created automatic procedure for code check pointing

Example of Special Assistance

“We have found NERSC staff extremely helpful in setting up

the computational environment, conducting calculations, and also

improving our software.”

Early Successes for INCITE 2007

• Tuning the FLASH code for memory use to correct an error condition

– “We could not have asked for better or more support than we got from the folks at NERSC, in helping us to get on the NERSC machines quickly, in giving the job special status, and in helping us meet the challenges of running a large job on Bassi.”

—Don Lamb• Tuning and debugging the

global tropospheric circulation analysis code– Adding multi-level

parallelism to bundle several associated parallel jobs

Science-Driven Systems

Science-Driven Systems

• Balanced and timely introduction of best new technology for complete computational systems (computing, storage, networking, analytics)

• Engage and work directly with vendors in addressing the SC requirements in their roadmaps

• Collaborate with DOE labs and other sites in technology evaluation and introduction

Computational System Strategy

• Balanced and timely introduction of the best new technologies for complete systems – Often the first and/or largest of its kind

• Computational systems address the widest breadth of DOE capability science – NERSC has two major computational systems in place

at a time – called NERSC-n• NERSC’s major computational systems arrive every three

to four years– The oldest-generation system is replaced with the latest-

generation system.– Modest-sized systems (NCSy) arrive between the major

systems as funding and technology allows• Typically focus on a subset of the workload

NERSC Systems 2007

ETHERNET10/100/1,000 Megabit

FC Disk

STKRobots

HPSS100 TB of cache disk

8 STK robots, 44,000 tape slots, max capacity 44 PB

PDSF~1,000 processors

~1.5 TF, 1.2 TB of Memory~300 TB of Shared Disk

Ratio = (0.8, 20)

Ratio = (RAM Bytes per Flop, Disk Bytes per Flop)

Testbeds and servers

Visualization and Post Processing Server- davinci64 Processors.4 TB Memory

60 Terabytes Disk

HPSS

HPSSNCS-b – Bassi

976 Power 5+ CPUsSSP5 - ~.8 Tflop/s

6.7 TF, 4 TB Memory70 TB disk

Ratio = (0.5, 9)

NERSC Global File System

300 TB shared usable disk

StorageFabric

OC 192 – 10,000 Mbps

IBM SPNERSC-3 – “Seaborg”

6,656 Processors (Peak 10 TFlop/s)SSP5 – .9 Tflop/s

7.8 Terabyte Memory55 Terabytes of Shared Disk

Ratio = (0.8,4.8)

10 Gigabit,Jumbo 10 Gigabit

Ethernet

NCS-a Cluster – “jacquard”650 CPU

Opteron/Infiniband 4X/12X 3.1 TF/ 1.2 TB memory

SSP - .41 Tflop/s30 TB Disk

Ratio = (.4,10)Cray XT4

NERSC-5 - “Franklin”SSP ~16 Tflop/s

Franklin arrives, January 16, 2007

Largest XTLargest XT--4 4 9,740 nodes with 19,480 CPUs (cores) 9,740 nodes with 19,480 CPUs (cores)

102 Node Cabinets, 16 102 Node Cabinets, 16 KWsKWs per cabinetper cabinet39.5 39.5 TBsTBs Aggregate MemoryAggregate Memory

16.1+ Tflop/s Sustained System Performance16.1+ Tflop/s Sustained System PerformanceSeaborg - ~.89

101.5 Tflop/s Theoretical System Peak Performance101.5 Tflop/s Theoretical System Peak PerformanceCray SeaStar2/3D Torus Interconnect (17x24x24)Cray SeaStar2/3D Torus Interconnect (17x24x24)

6.3 TB/s Bi-Section Bandwidth 7.6 GB/s peak bi-directional bandwidth per link

50 Nanosecond per link latency345 345 TBsTBs of Usable Shared Disk of Usable Shared Disk

Sixty 4 Gbps Fibre Channel Data ConnectionsSixty 4 Gbps Fibre Channel Data ConnectionsFour 10 Gbps Ethernet Network ConnectionsFour 10 Gbps Ethernet Network Connections

Sixteen 1 Gbps Ethernet Network ConnectionsSixteen 1 Gbps Ethernet Network Connections

Benjamin Franklin, one of America’s first scientists, performed ground breaking work in energy efficiency, electricity, materials, climate, ocean currents, transportation, health, medicine, acoustics and heat transfer.

““FranklinFranklin””

2006: NERSC Global Filesystem(NGF) in Full Production

• After thorough evaluation and testing phase in production

• Based on IBM GPFS• Seamless data access from all

of NERSC’s computational and analysis resources

• Single unified namespace makes it easier for users to manage their data across multiple system

• First production global filesystem spanning five platforms, three architectures, and four different vendors

NERSC Storage Roadmap

• PastLocal Disk

/scratch/home/tmp

HPSS

• NowLocal Disk

/scratch/home

NGF/project

HPSS

• FutureLocal Disk

/homeNGF-HPSS

/scratch/project

File System A

{Network

Machine A

/scratch

/home

File System B

Machine A

/scratch

/home

/project

NGFHPSS

{ Machine B

/scratch

/home/temp

File System B

Machine A

/scratch

/home/temp

File System A

Network

HPSS

{Network

/project/scratch/home

NGF & HPSS

Machine AMachine B

Science-Driven Analytics

What is the Analytics Program at NERSC?

• Analytics is the "science of analysis". • At NERSC, the Analytics Program is the

confluence of several key technologies:– Data management

• Data storage/retrieval/sharing/movement, data indexing/querying

– Data analysis and data mining • Compare datasets, find features within a dataset.

– Data visualization of data• Primary method of data exploration.

– Workflow management• Systematic approach to “data processing pipelines,”

especially those that use multiple distributed resources and automate scientific data processing activities

NERSC Analytics Services

• A dedicated interactive analysis platform that is well integrated with the rest of the Center

• Software applications, libraries, etc. that are the building blocks for Analytics solutions

• In-depth, collaborative help to science stakeholders to implement effective analytics solutions– There is no “one-size-fits-all” solution due to the diversity of

stakeholder problems• Serving shared licenses to NERSC users• Technology path-finding to stay abreast of latest

developments in the field, including interactions with the CS research community

Production Analytics –SNFactory

• New supernova data analysis and workflow visualization tools (Sunfalland SNwarehouse) have improved usability and situational awareness, and enabled faster and easier access to data for supernova scientists worldwide

• Advanced image processing (Fourier contour analysis) and machine learning techniques running on NERSC platforms have achieved a >40% decrease in human workload in nightly supernova search (>75% FTE)

Summary

• NERSC supports a diverse range of requirements and science

• NERSC systems, services and analytics are highly regarded and valuable to the DOE science community

• NERSC is helping to create the highly successful systems of the future